Towards Practical Mean Bounds for Small Samples

Authors: My Phan, Philip Thomas, Erik Learned-Miller

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In simulations, we show that for many distributions, the gain over Anderson s bound is substantial. 5. Simulations We perform simulations to compare our bounds to Hoeffding s inequality, Anderson s bound, Maurer and Pontil s, and Student-t s bound (Student, 1908), the latter being
Researcher Affiliation Academia My Phan 1 Philip S. Thomas 1 Erik Learned-Miller 1 1College of Information and Computer Sciences, University of Massachusetts, Amherst, MA, USA. Correspondence to: My Phan <myphan@cs.umass.edu>.
Pseudocode Yes Algorithm 1 Monte Carlo estimation of mα D+,T (x) where D+ = [0, 1]. This pseudocode uses 1-based array indexing.
Open Source Code Yes Code accompanying this paper is available at https://github.com/myphan9/small_sample_mean_bounds.
Open Datasets No We perform experiments on three distributions: beta(1, 5) (skewed right), uniform(0, 1) and beta(5, 1) (skewed left). Their PDFs are included in the supplementary material for reference. The paper uses synthetic data generated from these distributions, not pre-existing publicly available datasets that require a specific link or citation for access.
Dataset Splits No The paper conducts simulations by sampling from specified distributions (beta(1,5), uniform(0,1), beta(5,1)) for various sample sizes, but does not mention traditional train/validation/test dataset splits as it's not training a model on a fixed dataset.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, or memory specifications) used for running the simulations.
Software Dependencies No The paper describes algorithms and statistical methods, but does not specify any software dependencies or libraries with version numbers (e.g., Python, PyTorch, SciPy versions) used for implementation or simulation.
Experiment Setup Yes We use α = 0.05, D+ = [0, 1] and l = 10,000 Monte Carlo samples. We consider two functions T: 1. Anderson: T(x) = bα,Anderson ℓ (x), again with ℓ= u And. Because this T is linear in x, it can be computed with the linear program in Eq. 42. 2. l2 norm: T(x) = (Pn i=1 x2 i )/n.